Human in the Loop: Approval Steps for Agents

Human in the Loop means an AI agent pauses before a critical action and a real person must confirm the action before it is executed. In the automation platform n8n, this can be linked directly to individual tools of an AI agent: the agent decides it needs a particular tool, execution pauses, an approval request goes to a human, and only after their confirmation does the action continue, or it is cancelled if rejected. For companies that are skeptical about AI agents out of concern over uncontrolled actions, this is the key lever for combining automation and control. As of: July 2026.

Why approval steps build trust in AI agents

Many companies hesitate to use AI agents because, in the worst case, an agent might send an email, trigger an order, or delete data without anyone being able to stop it. This concern is justified, because an agent makes its decision based on probabilities, not on hundred-percent certainty. An approval step addresses exactly this: it separates the agent's decision from the actual execution. The agent may propose, a human must confirm. For companies still at the beginning of their AI automation journey, this is often the difference between cautious testing and the confidence to let agents run in production.

How tool approval works technically in n8n

According to the n8n documentation on Human-in-the-Loop for Tools approval can be linked directly to individual tools of an AI agent, not to the entire workflow. Setup runs in three steps: in the Tools panel of the agent node, an approval channel is selected and connected with the appropriate credentials, then the tools that require approval are attached to this review step. As soon as the agent decides to use a tool that requires approval, the workflow pauses and sends a request to the chosen channel. Using the variables $tool.name and $tool.parameters, the reviewer sees exactly which tool is being called with which inputs before approving or rejecting. On approval, the action continues with the values proposed by the agent; on rejection, it is cancelled and the agent receives information that the action was not executed. As approval channels, the documentation names, among others, Slack, Microsoft Teams, Discord, Telegram, Gmail, Microsoft Outlook, WhatsApp, Google Chat, and n8n's native chat interface.

The Wait node as the technical foundation

Approval steps in n8n build on the Wait node. This node pauses the execution of a workflow and stores the current state in the database until a resume condition is met. For approvals, the On Webhook Call mode is relevant: n8n generates a unique resume URL at runtime, through which the workflow is resumed once the response arrives. Alternatively, the On Form Submitted mode is available, in which a form with free-text, dropdown, or date fields is displayed and the workflow only continues after submission, optionally with an automatic time limit. According to the documentation, it should be noted that the resume URL changes on partial executions, which is why the triggering node must run together with the Wait node. In addition, time specifications always use n8n's server time, regardless of the configured time zone.

Practical example: sending emails and placing orders only after confirmation

A typical scenario: an agent processes incoming customer inquiries and, when needed, automatically sends a reply email or triggers a reorder. Instead of allowing the send directly, the email or ordering tool is linked to an approval step. The agent drafts the message, the request goes via Slack or Teams to a responsible person, who sees the text and recipient data and confirms or rejects it with a click. Only then does the email actually leave the system, or the order is triggered. This preserves the time savings of automation while the final decision remains with a human. This is exactly the core idea behind AI agents, as NordFlux sets them up for SMEs: you retain control over critical steps, even while the rest of the process runs automatically.

Limitations and what to keep in mind

An approval step does not replace clean error handling in the rest of the workflow, and it naturally slows down any process it affects, since an action waits until someone actually responds. If approval is not granted in time, the workflow should have a time limit and a defined fallback path instead of pausing indefinitely. In addition, the pattern is best suited for individual critical actions, not for having every single step of an agent manually rubber-stamped, or the benefit of automation is lost. Anyone unsure which actions should really require approval should draw up a short list of critical, hard-to-reverse actions before the technical implementation.

Frequently asked questions about approval steps for AI agents

What is the difference between per-tool approval and approval for the entire workflow?

With tool-based approval in n8n, only a single tool of the agent is linked to the review step; other actions continue without interruption. This makes it possible to specifically require approval only for risky steps such as sending emails or triggering orders, while research or read actions remain automatic.

How long can a workflow wait for approval?

The Wait node stores the workflow state in the database and waits until the resume URL is called or a set time limit is reached. For the form variant, an automatic time limit can be configured according to the documentation, after which the workflow either continues without approval or aborts, depending on the configuration.

Which channels are suitable for approval requests?

The n8n documentation names, among others, Slack, Microsoft Teams, Discord, Telegram, Gmail, Microsoft Outlook, WhatsApp, Google Chat, and n8n's native chat interface as approval channels. The most sensible channel is the one the responsible team already uses daily, so approvals don't get left unanswered.

Does an approval step slow down the entire automation?

Only the affected sub-step, not the entire process. All actions that are not linked to an approval continue to run automatically. It is important to deliberately limit approvals to a few critical actions so the benefit of automation is preserved.